test(job-array-single-wal-record): large-array HA soak scenario

Extends the HA soak harness with a new --array-soak mode: small (N=2)
and large (N=10000) array submission sub-rounds, each with one clean
kill/failover, asserting one WAL record per submission regardless of
array size, per-element addressability (status/cancel/completion)
spread across low/mid/high indices post-failover, and a real
millisecond-resolution replay-time bound comparing large-array vs
small-array kill-to-addressable latency.

waldump.c gains a WAL_JOB_ARRAY_SUBMIT case in wal_type_name() so the
scenario can grep an exact type name for its live WAL-record-count
delta check.

Went through one remediation round: the replay-time measurement
initially used whole-second resolution (collapsed to 0s/n-a on a real
run) and never asserted on the comparison, giving zero actual evidence
for the O(1)-replay-time requirement. Fixed with $EPOCHREALTIME
(millisecond resolution -- this sandbox's BusyBox date silently
mangles the more common `date +%s%3N` idiom) and a real bound
assertion that hard-fails the run on a genuine O(N) regression,
independently verified to have real teeth.

Live-verified 3 times with genuinely discriminating numbers each time;
final run: small=210ms, large=163ms -- large-array replay was not
slower, live evidence of the feature's O(1) claim.

Task 6.1 of job-array-single-wal-record (Requirements 4.3, 5.2, 5.3,
8.2, 8.3). Depends on and reuses task 6.0a's own soak verification.

.kiro/specs/job-array-single-wal-record/tasks.md

@@ -166,7 +166,7 @@
   - _Requirements: 3.1, 3.2, 4.1, 9.4_
   - _Boundary: Array Apply Expansion (`persist.c`'s replay-side job-count accounting) — discovered during task 6.1's soak (2026-07-15), inserted per explicit user decision; blocks 6.1/6.2/6.3, all of which require a cluster that survives a restart after an array submission_
 
-- [ ] 6.1 Large-array soak scenario
+- [x] 6.1 Large-array soak scenario
   - Extend the existing HA soak harness with a large-array submission scenario asserting: one WAL record and one commit barrier regardless of array size, failover replay time governed by the record's size rather than element count, and every array element individually addressable (status, cancel, completion) after a failover
   - Observable: a soak run reporting the one-record/one-barrier assertion, a replay-time measurement independent of array size, and successful per-element status checks post-failover
   - _Requirements: 4.3, 5.2, 5.3, 8.2, 8.3_
@@ -205,3 +205,4 @@
 - **Task 5.4**: Go/CGO receive path — `rawJobArrayToModels()` (pure-Go N-way expansion), `cJobArrayToModels()` (CGO adapter sibling to `cJobToModel`), `dispatchJobArray()` (dedup via the SAME `store.Jobs.Has()` the ordinary `Data_data_jobs` path writes through, one confirmation log only on ≥1 newly-dispatched element), new `case C.Data_data_jobArray:` in `handleDataset`. Elements route through `OnJobUpdated` (not `OnJobCreated`), matching the ordinary single-job push's own routing — verified consistent, not a divergence. CGO boundary needed zero changes — `C.xce_job_submit_t` and `C.Data_data_jobArray` were already exposed/generated from prior tasks. A structural TOCTOU exists between the `Has()` dedup check and the eventual store write (two separate lock acquisitions, no compare-and-set primitive) but is narrower than task 5.3's accepted C-side gap (no intervening async hop) and strictly safer than the pre-existing baseline (the ordinary `Data_data_jobs` path has NO dedup check at all today) — not fixed, judged acceptable. `cJobArrayToModels()` has zero direct unit coverage at the CGO-struct level, but this exactly mirrors the pre-existing, repo-wide precedent for `cJobToModel()` itself (also never unit-tested with a real C struct) — reviewer independently reconstructed a full `CGO_ENABLED=1` build+link+test run against the real generated headers to confirm correctness beyond a compile-only check. `store.ArrayJobs` (array-level rollup aggregate) deliberately left unpopulated — confirmed genuinely out of this spec's scope per requirements.md's Boundary Context, which explicitly defers whole-array bulk query/cancel to a future spec; zero code anywhere in the codebase populates it today, including this diff.
 - **Task 5.5**: Cross-daemon integration test, test-only (zero net production diff, confirmed on `request.c` and every task 5.1-5.4 file). Drives real `xce_encode()`/`xce_decode()` wire bytes through the REAL production dispatch chain on both daemons — `extract_data()` → `scheduler_req_handler()` → `sched_handle_dataset()` → `sched_on_job_array()` (C, via `#include "../src/handler.c"` in the test) and `go_push_handler()` → `handleDataset()` → `cJobArrayToModels()`/`dispatchJobArray()` (Go/CGO) — skipping only `vtask_entry()`'s ~10 lines of `static` glue, which the test mirrors verbatim (independently verified line-by-line accurate on the one behavior that mattered: `if (callnext)` treating any nonzero `extract_data()` return, including a negative errno, as "proceed"). Redelivery is tested via genuinely independent fresh decodes of separately-encoded bytes, not a cheaper decode-once-call-twice shortcut. Bug-catching capability independently proven twice (implementer, then reviewer): temporarily reverting task 5.3b's fix makes both the C and Go tests fail meaningfully; restoring returns both to green. Go/cgo toolchain constraint discovered: `import "C"` is disallowed in any `_test.go` file, forcing a two-file split (`wire_dispatch_support.go` for cgo-preamble helpers, `wire_dispatch_test.go` for the actual tests) — logged as bug-1440/1441/1442. One low-severity, non-blocking nit (not fixed): the C test file's own header comment misdescribes which assertion fails first under the RED-phase revert (claims the `callnext != 0` check; actually a downstream `>= 0` check) — cosmetic only, the test's real bug-catching property is independently confirmed accurate.
 - **Task 6.0a**: Fixed `replay_wal()`'s job-count tally (persist.c ~3902) to add a `WAL_JOB_ARRAY_SUBMIT` branch that tallies `apply_rc` (the genuine per-record element count `apply_committed_entry()`'s array case already returns) when `apply_rc > 0`. Independently traced the FULL return-value space of `restore_job_array()`/`apply_committed_entry()`'s array case (6 distinct cases: 3 negative pre-mutation failures, `-EEXIST`→`apply_rc=0` idempotent no-op, the task-3.1 bulk-insert partial-failure case (stays negative, correctly still NOT tallied — a pre-existing, out-of-boundary undercount that still surfaces as a detectable anomaly rather than masking one), and positive `created` on genuine success) and confirmed every positive `apply_rc` this fix tallies is PROVABLY equal to real jobmgr growth for that record — the fix cannot overcount, and therefore cannot weaken `persist_verify_apply_correctness()`'s ability to catch a genuine divergence. Verified end-to-end via a from-scratch live 3-node HA soak re-run (`--array-soak`, both by the implementer and independently by the reviewer): zero `APPLY_DIVERGENT` across all three of the original bug's trigger paths (restart-mid-small-array, standby election-path recovery fallback, restart-mid-N=10000-array), one-WAL-record/one-commit-barrier confirmed for both N=2 and N=10000, per-element post-failover status/cancel/completion confirmed. This soak run also directly satisfies task 6.1's own Observable — its results should be reused rather than re-run from scratch. **New non-blocking finding (pre-existing, NOT introduced by this task)**: `controller_unit_test`'s single outer `UNITY_BEGIN()` is silently reset by ~19 of its called test-group functions' OWN nested `UNITY_BEGIN()/UNITY_END()` calls, making the process exit code / meson `Ok:/Fail:` summary line unreliable for detecting a regression in any group running before the last such nested call — both the implementer and reviewer worked around this by grepping raw `testlog.txt` for `:FAIL:` lines directly rather than trusting the summary, and this should be the standard practice going forward for this binary until a dedicated follow-up fixes the harness itself.
+- **Task 6.1**: Extended `.claude/skills/xce-test-ha/run-ha-soak.sh`/`soak-verdict.sh` with a new `--array-soak` mode (small N=2 + large N=10000 sub-rounds, one clean kill/failover per sub-round — deliberately single-kill, not chaotic, matching this task's correctness-not-chaos scope; chaos-under-array-submission is task 6.2's separate job). `waldump.c` gained a `WAL_JOB_ARRAY_SUBMIT` case in `wal_type_name()` so the scenario can grep an exact type name for its live before/after WAL-record-count delta check. **Round 1 REJECTED**: the replay-time measurement used whole-second (`date +%s`) resolution — both sub-rounds' timings collapsed to `0s`/`n/a` on a real run — and, more fundamentally, nothing ever asserted on the comparison (`verdict_emit_arraysoak()` always returned PASS regardless of measured values), leaving Requirement 4.3 ("replay time governed by record size, not element count") with zero actual evidence and unable to catch a genuine O(N) regression. **Round 2 fixed**: switched to `$EPOCHREALTIME` (millisecond resolution; discovered and worked around a genuine environment trap — this sandbox's BusyBox `date` precedes GNU coreutils on `$PATH` and silently passes `%N`/`%3N` through as a literal string suffix rather than expanding or erroring, which would have poisoned any `date +%s%3N`-based approach silently); added `verdict_check_arraysoak_replay_bound()` with a real `bound = max(t_small*ratio_max, floor_ms) + margin_ms` assertion that `run_array_soak_round()` hard-`fail()`s on when exceeded — independently verified to have real teeth (both the implementer's and reviewer's own hand-picked regression cases correctly trip `BOUND_EXCEEDED`). Live-reproduced 3 times total (once round 1, twice round 2) with genuinely discriminating millisecond numbers each time (final: small=210ms, large=163ms — large array replay was NOT slower, live evidence of the feature's O(1) claim). Also fixed opportunistically: status/cancel/completion checks spread across low/mid/high indices (previously clustered at 1,2,3) matching the already-good basic-addressability spread; selftest extended 17→28 cases covering the new bound-check logic; `SKILL.md` documented. "One commit barrier" remains an inference from the WAL-record-count delta (not independently observed live) — defensible given task 2.1/2.3's unit-tested 1:1 append↔continuation coupling, but flagged as an inference, not a direct check, should a future task want to close that gap.